Problem Description

When the gateway restarts (manually via systemctl restart or crash recovery), sessions that were mid-execution are loaded but the agent doesn't continue. The agent "hangs" waiting for user input.

Scenario:

1. User sends message: "check the latest commits"
2. Agent starts executing tools (e.g., terminal, search_files)
3. Gateway restarts mid-execution
4. Session history is loaded successfully
5. Last message in history is a tool role response
6. Agent waits for new input instead of continuing execution
7. User must write "continue" or use /retry to proceed

Root Cause

When history ends on a tool role message (tool response), the agent doesn't automatically continue the execution loop. The agent expects either:

• User input to start fresh
• An assistant message to be the last message (completed turn)

But when execution is interrupted by gateway restart, the assistant message was never written — execution stopped after the tool response.

Expected Behavior

After gateway restart, if the loaded session history ends on a tool role message, the agent should automatically continue execution as if it never stopped. This is similar to how /retry works, but should happen automatically.

Workaround

Currently users must either:

1. Write "continue" or any message to trigger continuation
2. Use /retry command to re-send the last user message

Proposed Solution

In gateway/run.py, after loading session history, check if the last message is a tool role. If so, automatically trigger a continuation:

# After loading history in _handle_message or similar
history = self.session_store.load_transcript(session_entry.session_id)

if history and len(history) > 0:
    last_msg = history[-1]
    if last_msg.get("role") == "tool":
        # History ends on tool response - execution was interrupted
        # Auto-continue by triggering agent with empty continuation prompt
        logger.info("Session history ends on tool response, auto-continuing execution")
        # ... trigger agent loop continuation

Alternative Solutions

1. Add /continue command — explicit command to resume interrupted execution without re-sending the user message
2. Persist execution state — save that agent was mid-execution, restore on startup
3. Send system message on restart — gateway sends a "Gateway restarted, continue" system message to agent

Environment

• Hermes Agent v0.6.0 (2026.3.30)
• Gateway running via systemd
• Platform: Telegram
• Session persistence: enabled (default)

Related Code

• gateway/run.py — session loading logic around line ~1700
• SessionStore.load_transcript() — history loading
• _handle_message() — where history is loaded and agent is invoked

Impact

This affects all gateway users who restart their gateway (manually or due to updates/crashes). The interruption of mid-execution tasks is a common pain point for long-running operations.